Anthropic's $14B ARR vs. $5B Cumulative Revenue

This piece looks at revenue metrics for businesses where AI inference costs are baked directly into the cost of the product.

$14 Billion Annualized Revenue, Over $5 Billion Cumulative Revenue

Hi, this is Oswarld. Today I want to talk about which numbers matter when you’re building a go-to-market strategy.

On February 12, 2026, Anthropic announced its Series G funding round and disclosed a run-rate revenue of $14 billion. The post-money valuation was $380 billion. Run-rate revenue annualizes the company’s most recent revenue level over a full year — it’s not the same as revenue actually earned over the past 12 months.

On March 9, CFO Krishna Rao filed a declaration in a lawsuit against the U.S. Department of Defense stating that cumulative revenue since the launch of commercial service had surpassed $5 billion. Here, the figure refers to the total revenue accumulated up to that point.

Both numbers can be true at the same time. If recent revenue has been growing fast, run-rate revenue can exceed the historical cumulative total. Also, “over $5 billion” doesn’t mean exactly $5 billion. What I want to focus on today isn’t the size of these numbers, but what each metric actually explains — and what it doesn’t.

User count, recurring revenue, and profit each tell a different story

Which metrics matter most depends on how a service operates. Advertising services care about visits and frequency of use; subscription services care about recurring revenue and customer retention. One metric doesn’t fully substitute for another.

As smartphones and social media grew, daily active users (DAU)¹ and monthly active users (MAU) came into frequent use. They show how many people use a service and how often they come back. They’re also tied to how many opportunities exist to show ads, but user counts alone can’t tell you advertising revenue.

Even with the same user count, revenue differs if the depth of engagement, ad pricing, or paid-conversion rate differs. You also need to check how “active” is defined and how duplicate or fake accounts are handled. DAU and MAU are useful for gauging scale of use, but they aren’t metrics that explain profitability.

In subscription businesses like SaaS², monthly recurring revenue (MRR) and annual recurring revenue (ARR) are widely watched. Multiplying MRR by 12 expresses current recurring revenue at an annualized scale. This can differ in scope from “run-rate,” which converts all recent revenue, including one-time revenue, to an annual figure.

ARR helps you grasp how large the current subscription base is and how much it’s grown compared to before. But it was never a metric that guaranteed profit or actual annual revenue from the outset. Customer churn and upsells, discounts, and the cost of delivering the service all need to be checked separately. SaaS businesses also incur server and customer-support costs, and gross margins³ vary from company to company.

As inference costs and per-customer usage differences grow larger in AI products, these complementary metrics need to be examined more closely.

Three More Things to Check in AI Products

First, costs that scale with usage. Inference⁴ — the process by which an AI model generates an answer — consumes computing resources. The more a customer leans on complex code generation or long-document analysis, the higher the cost can climb.

Second, how much margin each customer actually leaves behind. Here’s a hypothetical to illustrate the point. Say there are two customers, each paying $2,000 a month. Customer A has a gross margin of 78%, Customer B has 31%. Both generate the same $24,000 in annual recurring revenue, but their monthly gross profit is $1,560 and $620, respectively. ARR doesn’t capture this gap at all. Since B’s gross margin is still positive, you can’t immediately call B an unprofitable customer — you’d need to look further at support costs, sales costs, and so on. This kind of per-customer profitability analysis is called unit economics⁵.

Third, how long a customer sticks around. A simple API integration is relatively easy to swap out. But a product deeply woven into a customer’s data, permissions, evaluation criteria, and operational workflows takes real time and money to migrate away from. Rather than assuming every customer can walk away easily just because the product is “AI,” you need to look at actual renewal rates and the real reasons behind churn.

These three factors also shape pricing and sales strategy. Two customers can generate identical revenue, yet leave the company with very different profit once you account for differences in usage cost and retention length. If you focus only on how fast you’re acquiring customers, you can miss the burden that shows up once usage starts to grow.

Reading Company Announcements and Cost Data Together

When reading Anthropic’s and OpenAI’s announcements, we need to keep the definitions of revenue metrics separate from the cost data.

Anthropic’s annualized revenue of $14 billion shows the revenue scale at the time of the announcement. The court declaration we looked at earlier stated that cumulative revenue had exceeded $5 billion, while spending on training and inference had surpassed $10 billion. What this data confirms is that revenue growth required massive capital injection alongside it. Since both figures are presented only as lower bounds, we can’t subtract one from the other to calculate an exact cumulative loss.

OpenAI CFO Sarah Friar also stated in a post dated January 18, 2026, that 2025 ARR had exceeded $20 billion. This, too, is a figure presented as an annualized metric. To know the actual revenue and profit for that year, we’d need to look separately at the full-period results and cost data.

Cost data also requires checking the scope. Ed Zitron reported that Anthropic’s AWS spending from January to September 2025 was about $2.66 billion, and estimated revenue for the same period at about $2.55 billion. The AWS spending figure is a payment amount confirmed with sources, while the revenue is an estimate backed out from public reporting. We shouldn’t take a payment figure that mixes training, inference, storage, and other costs, divide it against revenue, and read that directly as the service’s gross margin or accounting loss.

Adding cost and customer retention metrics to revenue

There are also proposals for what other metrics AI companies should add. I think the following two perspectives are especially useful.

The first is the cost incurred to generate that revenue. Even if ARR per employee rises, if high compensation or AI usage fees grow faster, it’s hard to say cost efficiency has actually improved. You need to check actual spending alongside headcount, not headcount alone.

GTM strategist Kyle Poyar has also proposed looking at ARR relative to money spent on people, rather than ARR per employee. For organizations with heavy AI usage costs, you can look at personnel costs and AI spending together. The costs being compared against annual ARR also need to be put on the same annual basis. This ratio itself doesn’t represent net margin.

Venture investor Tomasz Tunguz has proposed gross profit per token. Comparing estimates across 6 AI companies, he found a log-scale correlation coefficient of 0.70 with enterprise value, and 0.47 with token throughput. The sample is small, and revenue, cost, and usage figures are all estimates, so this can’t really be seen as proof of the investor’s judgment. Still, the underlying question—looking at usage together with the profit left over from that usage—is worth keeping in mind.

The second is how much of the first-year customer acquisition cost gets recovered. LTV/CAC⁶ compares the profit a customer will generate over their entire relationship with the company against the cost of acquiring them. But for products that launched recently, there isn’t enough data to estimate how long customers will actually stick around.

Because of this uncertainty, Poyar suggests focusing on first-year value rather than projected customer lifetime value. Rather than optimistically assuming returns several years out, the idea is to look at how much of the cost can be recovered from revenue within the first 12 months.

In practice, you can group customers who signed up around the same time (a cohort) and compare their first-year revenue against cost of service delivery, support costs, and customer acquisition cost. At the same time, you need to look at the outcomes customers actually achieved and whether they actually renewed. The profit that stays with the company and the value the customer gets from the product are related, but they aren’t the same calculation.

Oswarld’s Lens

One thing hasn’t changed: you have to pick metrics that fit your business model. For advertising businesses, usage frequency and ad revenue matter; for subscription businesses, recurring revenue and retention matter. For AI products, you need to connect these more precisely to per-customer usage costs.

From my experience building GTM strategy, metrics are both measurement tools and the incentives that determine what an organization prioritizes. If you evaluate a sales team purely on revenue, they’ll gravitate toward closing bigger deals. You need them looking at the cost side for high-usage customers too — otherwise you end up selling a lot without keeping enough of it.

I don’t think you can build lasting trust by simply touting large revenue figures. You need to explain what period the reported number covers, how it was calculated, and what the cost structure and retention rate look like. At the same time, you shouldn’t jump to accusing a company of inflating revenue just because different metrics show different things.

ARR isn’t rendered meaningless across every AI business. It’s still useful for tracking growth scale. But my view is that for products with heavy inference costs and wide variance in per-customer usage and margin, you need to pair revenue with cost and retention data.

Doing this requires connected data. If per-customer billing, model/infrastructure usage costs, and customer support records all live in separate silos, you can’t calculate profitability. The priority isn’t just naming a new metric — it’s making sure customer-level revenue and cost can actually be compared over the same time period.

Closing

User count shows scale of usage, and ARR shows scale of recurring revenue. To judge an AI product’s profitability, you need to add usage costs and customer retention rates to that picture.

When running an AI business or reading a company’s earnings, start by checking what the announced number is actually called and what period it covers. Then look at gross margin, cost per customer, and renewal rates together, and you’ll get a much more concrete picture of what’s really driving the growth.

References & Further Reading

• Anthropic, Series G funding announcement, 2/12/2026. Annualized revenue and post-money valuation.
• Krishna Rao, Court filing declaration, 3/9/2026. Cumulative revenue and training/inference spend in paragraph 8.
• Sarah Friar, A business that scales with the value of intelligence, 1/18/2026. OpenAI’s ARR announcement.
• Ed Zitron, This Is How Much Anthropic and Cursor Spend On Amazon Web Services, 10/20/2025. Reporting on AWS spending and the range of revenue estimates.
• Tomasz Tunguz, Gross Profit per Token, 12/30/2025, revised 1/2/2026. Analysis of estimates across 6 companies.
• Kyle Poyar, A proposal to rethink AI company metrics, 3/2026. A public post on ARR relative to headcount spend and first-year value.
• Stripe, Understanding MRR and ARR. The difference between recurring-revenue metrics and accounting revenue.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. DAU and MAU are the number of users active in a given day or month, respectively. You need to check what counts as “active” before comparing companies. ↩

2. SaaS (Software as a Service) is a model where software run by a provider is accessed over the internet, typically billed by subscription or usage. ↩

3. Gross margin is revenue minus cost of goods sold, divided by revenue. If revenue is 100 won and cost is 20 won, the margin is 80%. Not all sales, administrative, and R&D expenses are included in this calculation. ↩

4. Inference is the process of using a trained model to generate answers or predictions. ↩

5. Unit economics compares revenue and cost on a per-customer or per-transaction basis. You have to first decide which costs to include. ↩

6. LTV/CAC compares the profit a customer is expected to generate over their lifetime against the cost of acquiring that customer. The result shifts depending on assumptions about retention period and cost. ↩